To counter low recognition accuracy of unsupervised feature extraction methods and too many part of labels of supervised feature extraction methods,a semi-supervised feature extraction method based on few labeled samples is proposed,which can reduce data dimensionality and avoid small sample size problem to improve the recognition accuracy.The image dataset is performed by discrete cosine transform,the semi-supervised discriminant power based on frequency distribution is computed with labeled samples,and high semi-supervised discriminant power is seeked to extract representative features.The proposed method is combined with other feature extraction methods,and codnucted on different face databases.The results prove that the method is efficient and can obtain higher recognition accuracy than traditional methods with lower cost.